Helicone review

Helicone is a hosted AI gateway and LLM observability service, with a self-hosted server option. In a typical architecture, an application sends model traffic through Helicone's OpenAI-compatible proxy, or logs requests asynchronously. The gateway sits between application code and inference providers; the observability layer records request and response data for analysis.

The gateway provides access to more than 100 models through one API key, with routing strategies based on latency, cost or weights, automatic failover, and unified billing. The observability app tracks latency, token use, time to first token, errors and costs, with filters for model, user ID, session and custom properties. Helicone also offers agent tracing, prompt versioning and deployment, caching, and configurable rate limits. That makes it an infrastructure layer for applications already calling LLMs, rather than a model, client library or command-line tool.

The main integration surface is an OpenAI-compatible proxy: applications send model requests through Helicone, which captures request and response data. Teams can instead use asynchronous logging. The vendor describes a unified proxy endpoint and one API key for more than 100 models, plus direct integrations with OpenAI, Anthropic, Google Gemini, AWS Bedrock, Groq, DeepSeek, Together AI, Mistral and OpenRouter.

Framework and workflow integrations include LangChain, LlamaIndex, LangGraph, Vercel AI SDK, CrewAI, DSPy, Semantic Kernel, PostHog, n8n and Zapier. Prompt versions can be deployed through the gateway without changing application code, and the playground supports prompt, session and trace testing. The vendor does not name programming languages or a CLI in its integration details. Confirm SDK availability and language-specific setup in Helicone's docs before choosing an integration path. Its advertised surface is broad, but teams should validate that their existing request flow and framework fit the proxy or asynchronous logging model.

Hobby is free and includes 10,000 requests, 1 GB of storage, one seat, one organization, seven-day retention and ingestion of 10 logs per minute. That can cover a small project, but the low ingestion ceiling and short retention are constraints for production traffic.

Pro costs $79 per month and includes 10,000 free requests, then usage-based billing, unlimited seats, one organization, 1 GB of storage before usage-based charges, one-month retention and ingestion of 1,000 logs per minute. Team costs $799 per month with the same request and storage allowance structure, five organizations, three-month retention and 15,000 logs per minute. Both include a seven-day free trial. Enterprise has custom pricing, unlimited seats and organizations, forever retention, 30,000 logs per minute and on-prem deployment.

Usage-based request and storage rates are not published, and annual prices are not published. Ask Helicone for a cost estimate using expected traffic, storage and retention before routing production workloads through a paid plan.

Helicone is open source under the Apache 2.0 license. The vendor offers self-hosting with Docker Compose for local environments and production Helm charts for Kubernetes. Self-hosting gives organizations an option for greater data ownership, but it also means operating the required infrastructure; the vendor does not publish a data-export process, so confirm export formats and procedures before relying on the hosted service.

The OpenAI-compatible proxy and asynchronous logging provide integration surfaces that may be easier to replace than application-specific model calls. However, prompt versions, traces, dashboards and stored request data can create migration work. Teams should establish how to retrieve those records and map prompts and metadata to a replacement system. The license permits use of the source under Apache 2.0, but it does not eliminate the operational work of running or migrating the platform.

The managed service is described as SOC 2 and GDPR compliant and built on edge infrastructure with a Rust-based proxy. Those are useful deployment and compliance signals, not substitutes for checking the controls and contractual terms your organization requires. Match your expected volume against each plan's ingestion limit: Hobby lists 10 logs per minute, Pro 1,000, Team 15,000 and Enterprise 30,000. Confirm how request handling behaves at those limits and what happens during provider outages or gateway degradation.

Helicone advertises automatic failover, latency- and cost-based routing, caching backed by Redis or S3, and rate limits by request count, token use or dollar budget. The pricing page lists usage-based requests and storage beyond allowances, but not their rates. A Reddit thread reports that Helicone was acquired by Mintlify and that the project may be moving into maintenance mode, with security updates continuing but active feature development likely done. Treat that as a report to verify directly, especially if your production roadmap depends on new features or timely fixes.